Neural prior based reconstruction for robust autonomous navigation against various disturbances
Autonomous vehicles heavily rely on perception systems for urban navigation and environmental understanding.De-spite extensive researches about driving in favorable urban conditions,sensor failures and perception impairments under adverse weather and external interferences significantly impact the practical deployment of current autonomous driving systems.This paper proposed a neural prior-based autonomous driving information reconstruction algorithm for robust end-to-end naviga-tion.This algorithm densely stored scene geometry priors through implicit representation of driving scenarios and designed a reconstruction algorithm for perception based on the attention mechanism.In addition,it proposed a general framework to enhance the robustness of self-driving performance.Extensive experiments in the CARLA simulator demonstrate the generality and effectiveness of the proposed method,and the performance degradation rate of current self-driving models under external disturbances is reduced from 82.74％to 8.84％,which largely improves the driving performance of multiple existing self-driving models under external interferences.
万方数据
维普
